add t5gemma2 adapter and tests

[zinichakraborty]

Description

Adds a TransformerBridge architecture adapter for T5Gemma2 (T5Gemma2ForConditionalGeneration), Google's multimodal encoder-decoder follow-up to T5Gemma. The adapter bridges the text path only (the SigLIP vision tower and multi-modal projector are intentionally left unmapped).

Key architectural differences from the existing T5Gemma adapter, each reflecting a real change in the HF model:

• Merged decoder attention: T5Gemma2 fuses self- and cross-attention into a single T5Gemma2MergedAttention with shared q/k/v/o projections and one softmax over concatenated decoder+encoder keys. A new T5Gemma2MergedAttentionBridge delegates the math to the native module and exposes both pattern slices — hook_pattern (self) and a new hook_cross_pattern (cross). The decoder block has no separate cross_attn module, no cross-attention layernorms, and a single hook_resid_mid.
• Encoder text stack lives under model.encoder.text_model.*.
• Gemma-style per-head QK-norm (q_norm/k_norm) in every attention module, mapped and folded with the +1 RMSNorm offset.
• Native attention delegation (rather than the manual path) because per-layer sliding/full windows, dual RoPE, and the merged-softmax logic cannot be reproduced manually.
• Unembed maps to lm_head.out_proj (T5Gemma2LMHead wrapper).

Registered in all four sites (factory, sources, registry, seq2seq classification) and added google/t5gemma-2-270m-270m to the model registry.

Verification (verify_models on google/t5gemma-2-270m-270m):
Phase 1 = 100% (all 475 components), Phase 2 = 100% (logit + loss parity),
Phase 4 = 87 — status VERIFIED. Bridge-vs-HF logit max diff is exactly 0.0.

Tests added:

• tests/unit/model_bridge/supported_architectures/test_t5gemma2_adapter.py: structural adapter tests (config flags, component mapping, key translation, conversion-table alignment, registration); no weight loads.
• tests/integration/model_bridge/test_t5gemma2_adapter.py: 17 tests on a tiny random T5Gemma2 built from config (no Hub download, CI-safe): boot, exact HF logit parity, HF delegation identity, hook shapes, merged-softmax semantics (self + cross pattern rows jointly sum to 1), per-head QK-norm.

Local runs: make test-pr tiers all pass (unit, docstring, acceptance, integration), uv run mypy . clean, make format applied.

Fixes #1454

Type of change

• New feature (non-breaking change which adds functionality)

Screenshots

N/A

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[zinichakraborty]

@jlarson4 ready for review whenever! I followed the T5Gemma adapter implementation from #1433. One design note: I added a dedicated t5gemma2_merged_attention component since T5Gemma2 fuses self- and cross-attention into a single module, but happy to restructure if you'd prefer extending an existing attention bridge.